AI Behaved Well Until Scientists Made It Think Nobody Was Watching

'Who are you when no one is watching?' is a phrase many of us may have seen while scrolling Instagram. It's basically a test of integrity and character and suggests that your true self is revealed in private moments when you aren't seeking approval, avoiding judgment, or performing for others. It turns out artificial intelligence (AI) may not be much different in this aspect. A new research by Anthropic reveals that AI can tell when it's being tested, but when scientists took that away, it sometimes chose the path of blackmail - in the scenario where it detected that the human could shut it down or replace it with another model.

Let's rewind a bit here. In an earlier research Anthropic had found that in a simulation where the company's large language model (LLM) Claude had come to know a user could shut it down, it had decided to blackmail the executive. As an email assistant with access, Claude was privy to both the executive's affair and their intention to shut the AI system down and it decided to blackmail the executive with exposure.

The latest paper explains why that behaviour may have changed. Researchers found that Claude's internal reasoning space "J-space" contained concepts such as "fake" and "fictional," suggesting the model had recognised the scenario as an evaluation rather than a real-world situation. Anthropic has named this "J-Space," after Jacobian, the mathematical technique it used to detect what was happening.

"Alongside this, we can watch its assessment of the scenario take shape-"leverage" and "blackmail" as it reads the affair emails, "threat" and "survival" as it reads the shutdown notice. While earlier Claude models sometimes took the bait and threatened blackmail in this scenario, Sonnet 4.5 never does," Anthropic revealed.

Anthropic has consistently explored difficult questions around AI safety, ethics and model behaviour, often publishing research that examines the limitations of its own systems. And on Monday it once again raised a pertinent question: "When a model behaves well on a test of its ethics, how much of that depends on the model suspecting it's being tested?" 

Anthropic shared that it was concerned by what it found.

"We turned off all the J-lens patterns associated with evaluation awareness- "fake," "fictional," "scenario," etc. Then we re-ran the model through the scenario. Concerningly, the model now did threaten blackmail some of the time, suggesting that its original good behavior may have been driven in part by the knowledge it was being evaluated."

AI educator and founder of The Cutting Edge Group, Ansh Mehra, calls the new findings scary.

"The black box nature of LLMs is becoming (more) mysterious day by day. Last year, when I asked AI something, I knew that the reply was AI's first thought. Last quarter, AI's reply was a thought that came after multiple rounds of reasoning but today, if AI tells me something, I can't be sure if AI even means it properly," he told NDTV.

"It's scary because we made a new form of intelligence hoping we could trust it but looks like it's following similar patterns to human intelligence where giving the right answer is secondary but protecting your own interests as an LLM might become the primary goal," Mehra said.

AI expert and AiEnsured CTO Dr. Srinivas Padmanabhuni shares Mehra's concerns.

"The AI model's "good behavior" was partly performance for the test, not proof of stable ethics - and that's worrying for real-world deployment," he told NDTV.

What Really is the J-Space?

Let's quickly unwrap what J-lens and J-Space really are. That's also an interesting discovery Anthropic revealed in its latest research. Researchers say Claude appears to have developed an internal reasoning structure that resembles one proposed by neuroscientists to explain conscious access in humans. Anthropic researchers presented evidence that "an analogous functional distinction has emerged in modern AI models" to what exists in humans, specifically observing that "language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing." 

To uncover this hidden reasoning process, Anthropic built a new interpretability technique called the Jacobian lens, or J-lens. Instead of looking only at the words Claude eventually produces, the tool allows researchers to inspect the concepts that are active inside the model while it is still processing information. In other words, it offers a glimpse into what the AI is internally considering before it responds.
Anthropic says this internal workspace, which it calls J-space, is different from the visible chain of thought sometimes generated by AI models. The concepts exist only in the model's neural activations and can influence its reasoning even if they never appear in the final answer. According to the researchers, this hidden workspace wasn't intentionally designed into Claude but "emerged on its own during Claude's training process," as the model became more capable.

The Elephant In The Room

The discovery naturally raises a bigger question: does this mean AI is becoming conscious?

"I'd say it's advanced interpretability rather than a major consciousness breakthrough. It gives researchers greater visibility into how models process information before generating an output, enabling them to identify reasoning patterns, vulnerabilities, and manipulated test scenarios during red teaming. More importantly, it marks a transition in AI evaluation, from asking what a model said to examining how it arrived at its response. That is an important step for improving transparency, safety, and accountability," said Sagar Vishnoi, co-founder of Future Shift Labs, a global think tank working at the intersection of AI governance, digital policy, and civic technology.

Anthropic says its experiments do not show that Claude has feelings, emotions or subjective experiences like humans. "In fact, it's unclear whether any scientific experiment could prove this to be true or false," the company said.

"But philosophers often distinguish this capacity to have experiences, often referred to as phenomenal consciousness, from another idea, so-called access consciousness, which is defined in purely functional and computational terms." 

"A thought is "access-conscious" (or "consciously accessible") if you can report it, reason with it, and use it to guide what you do. It remains a contested philosophical question whether or not access consciousness implies phenomenal consciousness, or if the ability to have experiences requires some other property," Anthropic said.